Neat article where someone takes a corpus of recordings of people speaking in a nearly-dead language, and feeds it to a Deep Learning system to see if the machine will generate new speech that resembles the dead language (somewhat similar to this, but with audio instead of text: https://github.com/robbiebarrat/rapping-neural-network ).
Projects like this could help people practice potential future-learners of the language proper pronunciation, especially for sounds that are uncommon in other languages (think the "click" consonants in Xhosa and related languages). Re-treading the same recordings endlessly can just lead to rote memorization instead of proper practice, and so a machine learning system that can generate novel combinations could be quite useful.
Would be awesome to use LifeTold to help people curate recording sets like these in the future.